AI market research: checking a product launch or a new market before you commit
AI market research uses AI tools to settle the questions that come before a product launch or a move into a new market: how big the market is, who already sells into it and at what price, and whether customers want what you plan to offer. For a business that makes or sells physical goods, most of the evidence sits in public business and trade data, in competitors’ catalogues and in your own quotes and customer emails. AI reads far more of that material than a person can in a week. It also invents sources and numbers, so every figure it returns needs checking. This guide covers where AI helps, where it fails and how to run a small study.
What market research covers
Before a launch or a move into a new region, a business that makes or sells physical goods needs four answers:
- How many potential buyers there are, and how much they buy today.
- Who supplies them now, at what price and on what terms.
- What would make them switch, and what they would pay.
- What it takes to sell there: certification, labelling, distribution, freight and duties.
The first two answers can come from desk research. The third needs conversations with customers. The fourth usually needs a specialist, such as a customs broker or a regulatory consultant. AI for market research helps most with the desk research and least with the customer conversations.
How it is done by hand
Without AI, a smaller company researches a market with a few familiar methods:
- Searching the web and reading competitors’ websites one at a time.
- Asking distributors, sales agents and customers, often at trade shows.
- Buying an industry report, when one exists for the niche.
- Pulling government statistics into a spreadsheet.
Each step takes days, so the research is often cut short and the decision rests on a handful of conversations. AI shortens the desk research, which leaves more time for those conversations.
What AI changes
Research agents do the first pass
Several AI assistants include a research mode that plans a search, reads many web pages and writes a report with links. OpenAI introduced deep research in ChatGPT on February 2, 2025. It says the feature can “find, analyze, and synthesize hundreds of online sources,” and that each output comes “with clear citations.” Google says Gemini Deep Research can “automatically browse hundreds of websites” and turns your prompt into a research plan you can edit before it runs. Anthropic says the Research feature in Claude runs “multiple searches that build on each other” and returns “easy-to-check citations.”
These tools do a good first pass: listing competitors, collecting published prices and summarizing the rules for a market. OpenAI’s own announcement lists the limits. Deep research “can sometimes hallucinate facts in responses or make incorrect inferences” and “may struggle with distinguishing authoritative information from rumors.” Treat the report as a list of leads, and open every source before a number goes into a decision.
Public data sizes the market
Government statistics answer the sizing questions better than any model’s memory. AI helps you find the right table and set up the arithmetic in a spreadsheet you can check. Two classification systems connect the sources: NAICS, the North American industry codes, for your customers, and HS, the Harmonized System of customs codes, for your product.
- Count potential business customers. Statistics Canada’s Canadian Business Counts gives the number of business locations with employees by industry and employment size, for census metropolitan areas and census subdivisions. The US Census Bureau’s County Business Patterns gives the number of establishments, employment and payroll by industry and employment size.
- Measure trade in a product. Statistics Canada’s Canadian International Merchandise Trade web application shows imports at the HS6 or HS10 level and exports at the HS6 or HS8 level. The US Census Bureau publishes US import and export data by trading partner and commodity through USA Trade Online and an API.
- Find who imports it. The Canadian Importers Database, from Innovation, Science and Economic Development Canada, “provides lists of companies importing goods into Canada, by product, by city, and by country of origin,” based on Canada Border Services Agency data.
A language model can suggest likely NAICS and HS codes from a plain description of your customers and your product. Confirm product codes with a customs broker before you rely on them for duties.
Competitor research becomes a table
AI reads competitors’ catalogues, specification sheets, distributor listings and published price lists, and turns them into one comparison table: product, specifications, list price, pack size, lead time and warranty. It can repeat the check every month and flag what changed. Pricing intelligence covers ongoing competitor price monitoring, and cost-plus pricing and price optimization cover setting your own launch price. Respect each website’s terms of use, and keep a dated copy of every page you rely on.
Your own records hold demand evidence
The best evidence of unmet demand is often already in your systems: quote requests for items you do not carry, lost quotes with a reason, sales reps’ notes and customer emails. A language model can read a year of them, group the requests by product and need, and count each group. Remove personal details first. The Office of the Privacy Commissioner of Canada’s principles for generative AI advise organizations to “use anonymized or de-identified information within prompts to a generative AI system rather than personal information” where possible and reasonable.
Simulated respondents need a check
Some tools ask a language model to answer a survey as if it were a customer. The research so far calls for caution. In a Harvard Business School working paper revised April 30, 2026, James Brand, Ayelet Israeli and Donald Ngwe tested this. They found that willingness-to-pay estimates from language models “are sometimes comparable to estimates from human studies, but are often inaccurate and in some cases wrong-signed.” Fine-tuning a model on earlier surveys improved its answers for existing and new features within the same product category. The authors did not find the same improvement “for new product categories or for differences between customer segments,” which is often what a launch needs to know. They conclude that language models fit market research as a supplement to human studies.
Three examples
Each example shows the part AI does well and the part that still needs a person.
| Business | Question | Where AI helps | What still needs a person |
|---|---|---|---|
| A machine shop planning its own product line | How many shops could buy it, and what do competing products cost? | Counting machine shops by region, and a table of competing products and published prices | Calls with 10 to 20 shop owners about price and switching |
| An Ontario distributor opening in Quebec | Which customers are there, and who serves them today? | Business counts by census metropolitan area, and lists of competitors’ branches and product lines | Customer visits, and a review of language and labelling requirements by someone qualified |
| A food producer exporting to the US | How much does the US import, from where, and at what price? | US import data by product and trading partner, and competing products on retailer websites | Labelling and food safety rules checked with the regulator or a broker, and meetings with buyers |
A closer look at the machine shop
Say a machine shop wants to sell a line of its own workholding products to other shops. A research agent lists competing products and their published prices in an afternoon, and someone opens each source to check it. Statistics Canada’s business counts give the number of machine shops by census metropolitan area, which sizes the market within a day’s drive. The shop’s own quote history shows how often customers asked for workholding it did not make. Then the owner calls 15 of the shops on the list and asks what they use today, what it costs them and what would make them switch. The owner makes the decision after those calls, using the desk research to choose the shops and the questions.
AI market research tools
AI market research tools fall into four groups. This guide does not rank products, and features change often, so test each one on your own question.
| Kind of tool | What it does | Worth knowing |
|---|---|---|
| Research modes in AI assistants | Plans a search, reads many sources and writes a cited report | A fast first pass, and every citation needs opening |
| Survey and interview platforms with AI analysis | Drafts questions, transcribes interviews and groups the answers into themes | The answers come from real people, which is what a launch decision needs |
| Public data portals | Statistics Canada, the US Census Bureau and the Canadian Importers Database | Official figures. AI helps find the right table and do the arithmetic |
| A setup on your own data | Reads your quotes, lost orders and customer emails alongside public data | Keeps confidential records under your control |
Ask any vendor where your prompts and uploaded files are stored, whether they are used to train its models, and how it charges.
The data AI market research needs
| Source | Examples | What it answers |
|---|---|---|
| Public statistics | Business counts, trade data, building permits, census tables | Market size and where the buyers are |
| Competitor sources | Websites, catalogues, price lists, distributor listings, trade show exhibitor lists | Who sells what, and at what price |
| Your own records | Quotes, lost orders, CRM notes, customer emails, sales by region | Unmet demand, and why you win or lose |
| Customer conversations | Interview notes and survey answers collected by people | What buyers would switch for, and what they would pay |
Sales history adds a fifth source after the product reaches the market. AI demand forecasting covers turning early orders into a forecast.
How to start small
- Write the decision and the question in one sentence each. For example: should we open a branch in Montréal, and how many of our target customers are within 50 km of it?
- Run a research agent for the first pass, and save its report with every link.
- Open every source the report cites, and replace each number with the figure from the original table.
- Pull the public data tables yourself, and redo the arithmetic in a spreadsheet.
- Read your own quotes and emails for evidence of demand, with personal details removed.
- Talk to customers. Use AI to prepare the questions and summarize the notes.
- Write a one-page answer with each claim linked to its source.
Check a market before you commit to it
Tell Derik which product or region you are weighing and what you already know about it. He will tell you which questions your data and public sources can answer.
Start a conversationRisks and limits
- Invented sources and numbers. In Zhang v. Chen, 2024 BCSC 285, a lawyer filed two cases that ChatGPT had invented, and the Supreme Court of British Columbia ordered that the extra costs they caused be “borne personally” by her. A market study has no opposing counsel to catch a fake citation, so the checking has to be part of the method.
- Missing and dated information. Research agents read what is public and indexed. Paid industry reports, private distributor price lists and last month’s price changes are often missing, so a report can read as complete when it is not.
- Confidential plans. Check each tool’s data terms before you enter a launch plan, and prefer an account your company controls. Shadow AI covers setting a policy for staff, and private AI for business covers keeping company data under your control.
- Thin coverage of niche markets. Research agents favour what is written about online. A niche industrial market with few public sources can come back thin, or filled out with guesses.
- Simulated answers. The Harvard study above found the weakest results for new product categories, which is the usual situation for a launch.
- Codes and rules with legal weight. HS codes, labelling rules and certifications need confirmation from a qualified person.
How ThriveAI helps
ThriveAI is an AI engineering company in Ottawa that builds private AI systems on a company’s own data, for businesses that make, move or sell physical goods. For market research, that means setting public data beside what your quotes, lost orders and customer emails already show, with every figure linked to its source so someone can check it before a decision. Every answer shows where it came from.
ThriveAI’s systems read your ERP and accounting system as they are, and every connection only reads data. They are designed to keep each client’s data on its own server in Canada. You choose a model on that server or a hosted model under a written zero data retention agreement, and a hosted model may process requests outside Canada. Derik Lawlis, the founder, leads every project and stays close to the build. About ThriveAI covers the company.